Enhanced Lightweight YOLO Model for Efficient Vehicle Detection in Satellite Imagery

Vehicle detection in satellite images is a challenging task due to the variability in scale and resolution, complex background, and variability in object appearance. One-stage detection models are currently state-of-the-art in object detection due to their faster detection times. However, these models have complex architectures that require powerful processing units to train while generating a large number of parameters and achieving slow detection speed on embedded devices. To solve these problems, this work proposes an enhanced lightweight object detection model based on the YOLOv4 Tiny model. The proposed model incorporates multiple modifications, including integrating a Mix-efficient layer aggregation network within its backbone network to optimize efficiency by reducing parameter generation. Additionally, an improved small efficient layer aggregation network is adopted in the modified path aggregation network to enhance feature extraction across various scales. Finally, the proposed model incorporates the Swish function and an extra YOLO head for detection. The experimental results evaluated on the VEDAI dataset demonstrated that the proposed model achieved a higher mean average precision value and generated the smallest model size compared to the other lightweight models. Moreover, the proposed model achieved real-time performance on the NVIDIA Jetson Nano. These findings demonstrate that the proposed model offers the best trade-offs in terms of detection accuracy, model size, and detection time, making it highly suitable for deployment on embedded devices with limited capacity.